yomiyasu is an agent skill that rewrites AI-generated Japanese to sound like a person wrote it, with 7 rules and a built-in linter

Anyone who has published a Japanese article that passed through a large language model knows what AI prose reads like: inanimate subjects acting on things, long lists of bullet points, bold phrases every few lines, and figures of speech that mean nothing in context. yomiyasu, an open-source agent skill at nanaism/yomiyasu from the Tokyo optimisation company ALGO ARTIS, rewrites that output into prose a Japanese reader recognises as human, using 7 rules a developer can read and a Python linter that scores the result.

The MIT-licensed project reached 1,359 stars and 29 forks on 2026-10-04, four days after its first public commit on 2026-09-30. It installs into Claude Code, OpenAI Codex, Cursor and Gemini, and the author published a long essay about the design on Zenn.

What the skill actually does

The project describes 7 transformation rules the skill applies to AI-written Japanese. Restore full SVOCM structure, so each sentence names who is doing what to what. Make the actor and the request explicit, so a specification says "the system can" and a reader instruction says "please do". Dismantle non-living subjects, so concepts and tools no longer act as if they have intent. Replace figurative verbs such as "break", "knock down", or "melt" with the direct operation or state change they stand for. Cut unnecessary opening phrases and rhetorical double negatives, rewriting them as plain positive sentences. Preserve the original's technical constraints and numbers without adding new ones. Keep sentences at 30 to 45 characters on average with no more than 0 to 2 commas, and strip emoji, trailing colons and stray half-width spaces.

Three domain packs apply the rules differently: tech for technical articles, keeping steps intact; business for specifications and proposals, naming the responsible party; essay for personal blog posts, preserving plain feeling without inflating it into a lesson.

Two Python scripts anyone can run

The repository ships two scripts in scripts/, both using only the Python standard library. yomiyasu_lint.py scores a Markdown file for AI tells: bold frequency, bullet-list ratio, figurative verbs, emoji, trailing colons, repeated sentence endings, and the GitHub Markdown rendering fault where bold markers next to Japanese punctuation fail to render. It returns exit code 1 in strict mode, so a CI check or a git hook can fail a draft before it ships. yomiyasu_diff.py compares the original and the rewrite, flagging unintended meaning changes, invented information, and shifts in the stance of a sentence from suggestion to rule.

The repository includes a 50-fixture regression test suite and a comparison benchmark that uses open-licensed Japanese government documents and OSS design notes as inputs. Each test compares a plain LLM output against the yomiyasu-processed version, with the specific improvements written out.

Where the rules came from

The project credits a list of Japanese-language community research in its acknowledgements. A Speaker Deck talk by @nasuvitz argued that AI prose fails at the syntactic level before the vocabulary one, which the author cites as the basis for the SVOCM rule. A Qiita study by @nyosegawa measured bold and bullet-list inflation across 70,000 Qiita articles after AI tools became common, which the author uses to justify the format caps. Several blog posts analysing specific figurative verbs and rhetorical structures informed rules 4 and 5.

A team publishing Japanese content through a model now has a measurable standard instead of a subjective judgement. The author also notes that running yomiyasu alongside other Japanese-text correction skills can produce conflicting instructions, so other similar skills should be disabled before use.

Source

• Primary: nanaism/yomiyasu repository, GitHub
• Related: AI臭い日本語を脱臭するAgent Skill『yomiyasu』を作った話, Zenn